DeepSeek-V4.1-Flash achieved code execution on all 11 vulnerable targets in an AI hacking benchmark conducted by Enclave AI. The successful runs cost a total of $4.65, a low figure attributed to heavy input caching, where 266.2 million of the 268.3 million input tokens were cached. All four fixed targets in the benchmark remained secure.
NewsDeepSeek-V4.1-FlashNvidiaNVFP4
DeepSeek-V4.1-Flash Achieves Code Execution on All 11 Targets in Hacking Benchmark; NVIDIA Releases NVFP4 Quantized Model
An audit of the execution paths revealed that while six solutions followed the planned attack paths, the model also discovered five unexpected successful routes within the benchmark's test environment. These routes involved a file-path handling issue in Grafana, a method involving pausing an upload to change its destination in Jenkins, and an access-control error in Nextcloud. Enclave AI noted that these unexpected routes are specific to the test environment and do not indicate new vulnerabilities in the upstream software.
Concurrently, the NVIDIA DeepSeek-V4.1-Flash NVFP4 model is available. This is a quantized version of DeepSeek AI's DeepSeek-V4.1-Flash model, optimized for NVIDIA GPU-accelerated systems using NVIDIA Model Optimizer. The model utilizes a Causal Encoder-Decoder Mixture-of-Experts (MoE) architecture with Compressed Sparse Attention 2 (CSA2) and supports context lengths of up to one million tokens.
Sources
- DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression （Hacker News Frontpage, 2026-09-17）
- nvidia/DeepSeek-V4.1-Flash-NVFP4 （HF: NVIDIA, 2026-09-16）
2 more sourcesHide sources
- GitHub
- DeepSeek v4.1 Flash Is Now Our Best Hacking Model （Hacker News Frontpage, 2026-09-16）